Reproduce “ReF-LDM: A Latent Diffusion Model for Reference-based Face Image Restoration” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chi-Wei Hsiao, Yu-Lun Liu, Cheng-Kun Yang, Sheng-Po Kuo, Kevin Jou, Chia-Ping Chen
Researchers from MediaTek and National Yang Ming Chiao Tung University developed ReF-LDM, a Latent Diffusion Model for reference-based face image restoration, that leverages a novel CacheKV mechanism to efficiently integrate multiple non-aligned reference images. The model demonstrates superior identity preservation with high Identity Similarity scores and improved computational efficiency, overcoming limitations of prior methods.
Key insights
- Non-spatially aligned reference features can be efficiently integrated into Latent Diffusion Models by pre-caching Key (K) and Value (V) tokens from a single U-Net pass on reference images, then concatenating these with the KVs computed from the noisy latent in subsequent denoising steps (CacheKV).
- Applying an identity loss in diffusion models requires careful consideration of noise levels; scaling the loss by sqrt(alpha_t_bar) prevents quality degradation while effectively supervising identity preservation, especially at early, noisy timesteps.
- Leveraging multiple reference images significantly enhances identity preservation in face restoration, and a mechanism that can implicitly learn correspondences without explicit landmark detection is essential for robustness against severe degradation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.